我希望我能在改进椭圆拟合方法方面得到一些帮助。我正在考虑尝试使用RANSAC风格的方法,但我不确定它是否是正确的方向。任何有关我应该开始进入的方向的帮助将非常感激,即使它只是一个改进在我的边缘找到。
我一直在研究这个问题,而且我没有取得很大进展。我认为主要的问题是图像的质量,但我只能按照我现在的工作。
我正在测试的当前方法是在图像上使用边缘检测,然后尝试在我找到的边缘周围拟合椭圆。下面的图片将突出我的主要问题,即我的方法对噪音的处理非常差。
原始图片: http://i.imgur.com/usygfXw.jpg
Canny Edge检测后: http://i.imgur.com/K7XDcVL.png
椭圆拟合后: http://i.imgur.com/bN0lNIq.jpg
以下是我使用的代码。对于Canny边缘检测,我发现了一些值,现在静态使用它们。它是从网上获取的代码,然后我修改了,现在有点hacky抱歉。
#!/usr/bin/python
import cv2
import numpy as np
import sys
from numpy.linalg import eig, inv
# param is the result of canny edge detection
def process_image(img):
# for every pixel in the image:
for (x,y), intensity in np.ndenumerate(img):
# if the pixel is part of an edge:
if intensity == 255:
# determine if the edge is similar to an ellipse
ellipse_test(img, x, y)
def ellipse_test(img, i, j):
#poor coding practice but what I'm doing for now
global output, image
i_array = []
j_array = []
# flood fill i,j while storing all unique i,j values in arrays
flood_fill(img, i, j, i_array, j_array)
i_array = np.array(i_array)
j_array = np.array(j_array)
if i_array.size >= 10:
#put those values in a numpy array
#which can have an ellipse fit around it
array = []
for i, elm in enumerate(i_array):
array.append([int(j_array[i]), int(i_array[i])])
array = np.array([array])
ellp = cv2.fitEllipse(array)
cv2.ellipse(image, ellp, (0,0,0))
cv2.ellipse(output, ellp, (0,0,0))
def flood_fill(img, i, j, i_array, j_array):
if img[i][j] != 255:
return
# store i,j values
i_array.append(float(i))
j_array.append(float(j))
# mark i,j as 'visited'
img[i][j] = 250
# flood_fill adjacent and diagonal pixels
(i_max, j_max) = img.shape
if i - 1 > 0 and j - 1 > 0:
flood_fill(img, i - 1, j - 1, i_array, j_array)
if j - 1 > 0:
flood_fill(img, i, j - 1, i_array, j_array)
if i - 1 > 0:
flood_fill(img, i - 1, j, i_array, j_array)
if i + 1 < i_max and j + 1 < j_max:
flood_fill(img, i + 1, j + 1, i_array, j_array)
if j + 1 < j_max:
flood_fill(img, i, j + 1, i_array, j_array)
if i + 1 < i_max:
flood_fill(img, i + 1, j, i_array, j_array)
if i + 1 < i_max and j - 1 > 0:
flood_fill(img, i + 1, j - 1, i_array, j_array)
if i - 1 > 0 and j + 1 < j_max:
flood_fill(img, i - 1, j + 1, i_array, j_array)
image = cv2.imread(sys.argv[1], 0)
canny_result = cv2.GaussianBlur(image, (3,3), 0)
canny_result = cv2.Canny(canny_result, 107, 208,
apertureSize=3, L2gradient=False)
#output is a blank images which the ellipses are drawn on
output = np.zeros(image.shape, np.uint8)
output[:] = [255]
cv2.waitKey(0)
cv2.namedWindow("Canny result:", cv2.WINDOW_NORMAL)
cv2.imshow('Canny result:', canny_result)
print "Press any key to find the edges"
cv2.waitKey(0)
print "Now finding ellipses"
process_image(canny_result)
print "Ellipses found!"
cv2.namedWindow("Original image:", cv2.WINDOW_NORMAL)
cv2.imshow('Original image:', image)
cv2.namedWindow("Output image:", cv2.WINDOW_NORMAL)
cv2.imshow("Output image:", output)
cv2.waitKey(0)
答案 0 :(得分:2)
以下是我尝试过的内容,我使用dilate
和scipy.ndimage
来完成某些过程:
import cv2
import numpy as np
image = cv2.imread("ellipse.jpg", 0)
bimage = cv2.GaussianBlur(image, (3, 3), 0)
edge_image = cv2.Canny(bimage, 107, 208,
apertureSize=3, L2gradient=False)
img2 = cv2.dilate(edge_image, np.ones((3, 3)), iterations=3)
dis_image = cv2.cvtColor(img2, cv2.COLOR_GRAY2BGR)
import scipy.ndimage as ndimage
labels, count = ndimage.label(img2)
for lab, idx in enumerate(ndimage.find_objects(labels.astype(int)), 1):
sy = idx[0].start
sx = idx[1].start
y, x = np.where(labels[idx] == lab)
ellp = cv2.fitEllipse(np.column_stack((x+sx, y+sy)))
cv2.ellipse(dis_image, ellp, (0, 0, 255))
这是输出: